Jean-Pascal Pfister: Bayesian synapses

Universität Bern

Synapses are highly dynamical elements and operate on a wide range of time constants, from milliseconds to hours or even days. However, the computational principles that govern synaptic dynamics at those time scales are not clear. In this talk, I will argue that nonlinear Bayesian filtering - which aims at continuously estimating a changing latent variable from a stream of noisy observations -  acts as a unifying computational principle at different time scales. At the fast-time scale, we can view short-term synaptic plasticity as performing the optimal estimation of the presynaptic membrane potential [1]. At the slower time scale, long-term synaptic plasticity can be viewed as performing the optimal estimation of the target weights [2,3]. For spiking neurons, this Bayesian filtering perspective on long-term plasticity predicts several observed experimental findings such as Spike-Timing Dependent Plasticity as well as heterosynaptic plasticity and also offers novel testable predictions [3]. Taken together those results support the idea that synaptic plasticity can be meaningfully described from a normative Bayesian perspective.

 

[1] J.-P Pfister, P. Dayan and M. Lengyel. “Synapses with Short-Term Plasticity Are Optimal Estimators of Presynaptic Membrane Potentials.” Nature Neuroscience, 13(10), 2010: 1271–75.

[2] L. Aitchison, J. Jegminat, J.A. Menendez, J.-P. Pfister, A. Pouget and P. E. Latham. “Synaptic Plasticity as Bayesian Inference.” Nature Neuroscience 24(4), 2021: 565–71.

[3] J. Jegminat, S.C. Surace, and J.-P. Pfister. “Learning as Filtering: Implications for Spike-Based Plasticity.” PLOS Computational Biology 18(2), 2022: e1009721.

 

Guests are welcome!

 

Organized by

Tilo Schwalger / Margret Franke